Preprocessing Flashcards
1
Q
How are the terms session, run, volume, slice, voxel, in-plane resolution, field of view, slice thickness related?
A
- one session consists of multiple runs
- one run consists of multiple volumes
- one volume consists of multiple slices
- one slice consists of multiple voxels
- one slice is characterized by matrix size (amount of voxels) and in-plane resolution
- field of view = matrix size * in-plane resolution
- voxel size = in-plane resolution and slice thickness
2
Q
What are the 6 steps of preprocessing
A
- slice timing
- realignment
- coregistration
- segmentation
- normalization
- smoothing
3
Q
slice timing
A
- different slices are acquired during different time points
- measured signal of each voxel is Fourier-transformed, resulting in a frequency representation of the signal from which it can be reconstructed
- Fourier-transformed signal is phase-shifted to reference slice and back-transformed into signal space
- different reference slices give different statistical results
4
Q
realignment - why?
A
- head motion shifts measured signal between voxels -> failure to detect local activations (reduced sensitivity)
- motion can be correlated with experimental paradigm -> findings of spurious activations (reduced specificity)
5
Q
realignment - how?
A
- estimation/registration: determine rigid-body transformation from each acquired image to first (or mean) scan (3 translations and 3 rotations, reference volume - volume to realign = min)
- reslicing/resampling: apply estimated transformation to correct whole series of scans
6
Q
coregistration
A
- structural MRI needs to be in alignment with functional MRI before mapping to standard space (-> normalization)
- allows more accurate anatomical localization of activations
- practically, using the same algorithm as used for realignment
- using mean image from realignment
7
Q
segmentation - why?
A
- differentiation of structural MR image into tissue types can improve mapping to standard space (-> normalization)
- in SPM, segmentation and normalization are actually
performed using one model (“unified segmentation”) - brain tissue types: gray matter, white matter, CSF, meninges, skull, air
8
Q
segmentation - what?
A
- tissue probability maps
- normalization parameters used for warping the subject to standard space
9
Q
normalization - why?
A
- substantial inter-individual differences in brain anatomy
- increase sensitivity in analyses with multiple subjects by matching them to a standard brain -> anatomical template
- make results from different studies comparable by bringing them into a standard coordinate system -> MNI space
10
Q
normalization - how?
A
- linear registration: adjustment of global differences using an affine transformation with 4*3 = 12 parameters (translations, rotations, zooms, sheers)
- non-linear registration: adjustment of local differences using deformation fields based on smooth basis functions
11
Q
circular relationshipt between segmentation and normalization
A
- knowing which tissue type a voxel belongs to helps normalization
- knowing where a voxel is in standard space helps segmentation
solution: build generative model which accounts for both
- model how voxel intensities result from mixture of tissue type distributions
- model how tissue types have to be spatially deformed to match those of template
12
Q
normalization - best approach
A
- calculate mean functional image during realignment
- coregister structural image to mean functional image
- normalize coregistered mean functional image to anatomical template
- use resulting normalization parameters to normalize functional scans
13
Q
smoothing - why?
A
- residual anatomical mismatch, even after normalization
- reduce measurement artifacts, increase signal-to-noise ratio
- required by random field theory (RFT) -> statistical analysis
14
Q
smoothing - how?
A
- convolution with 3D Gaussian kernel defined by its full width at half maximum (FWHM)
- FWHM is directly related to standard deviation of a (multivariate) normal distribution (FWHM = 2 * sqrt(2 * ln(2 * sd))
- after smoothing, each voxel effectively becomes the result of applying a weighted region of interest
15
Q
When to normalize/smooth?
A
- to ensure that assumptions of random field theory (RFT) are met, use at least 2 x [normalized voxel size] as smoothing FWHM
- when using univariate approaches (e.g. voxel wise GLM), normalize and smooth before statistical analysis
- when using multivariate approaches (multi voxel pattern analysis), normalize and smooth after statistical analysis
